Dialogue of AI is throughout us, however in my expertise, sensible steerage rooted in particular use circumstances is surprisingly uncommon. After spending months deep within the weeds of an enormous documentation migration with AI as my assistant, I’ve discovered some hard-won classes that I believe others may benefit from.
In case you work in content material engineering, technical documentation, or are merely interested in how AI holds up in a posh, real-world venture, right here’s my tackle what labored and what didn’t.
Undertaking Context
I’m a DITA Info Architect on the Info Expertise crew at Splunk. DITA, brief for Darwin Info Typing Structure, is an open, XML-based commonplace for structuring and managing technical content material.
We lately wrapped up the migration of three massive documentation websites right into a single assist portal, powered by a DITA-based part content material administration system (CCMS). The timeline was tight, and practically all the sources have been inner. The migrations have been advanced and vital to the enterprise, requiring cautious planning and execution.
I initially deliberate solely to assist the migration of the smaller, unversioned website. When that went nicely, I used to be requested to guide the a lot bigger second migration. (The third website was dealt with by one other crew.) Collectively, these two migrations meant grappling with roughly 30,000 HTML information, two very totally different website architectures, and the problem of customizing an current Python migration script to suit the content material at hand, whereas additionally placing processes in place for writers to assessment and clear up their content material.
I need to be clear that AI didn’t full this venture for me. It enabled me to work sooner and extra effectively, although solely whereas I did the planning, architecting, and troubleshooting. Used successfully, AI grew to become an influence instrument that dramatically sped up supply, however it by no means changed the necessity for experience or oversight.
All through this venture, I used the then-current GPT-4 fashions by way of an inner Cisco chat-based deployment. Today, I work extra in editor-based instruments resembling GitHub Copilot. Nonetheless, the teachings I discovered ought to apply to the current (mid-2025) cutting-edge, with just a few caveats that I point out the place related.
How I used AI successfully
Prompting
One lesson I discovered early on was to deal with prompts the way in which I method technical documentation: clear, constant, and complete. Earlier than consulting the AI, I’d sketch out what wanted to occur, then break it down into granular steps and write a immediate that left as little to the creativeness as potential.
If I wasn’t positive in regards to the answer, I’d use the AI as a brainstorming associate first, then comply with up with a exact immediate for implementation.
Iterative improvement
The migration automation wasn’t a single script however grew to become a set of Python instruments that crawl navigation bushes, fetch HTML, convert to DITA XML, break up subjects into smaller models, map content material, and deal with model diffs. Every script began small, then grew as I layered in options.
I rapidly discovered that asking AI to rewrite a big script suddenly was a recipe for bugs and confusion. As a substitute, I added performance in small, well-defined increments. Every characteristic or repair bought its personal immediate and its personal GitLab commit. This made it straightforward to roll again when one thing went sideways and to trace precisely what every change achieved.
Debugging
Even with good prompts, AI-generated code hardly ever labored completely on the primary attempt – particularly because the scripts grew in dimension. My best debugging instrument was print statements. When the output wasn’t what I anticipated, I’d sprinkle print statements all through the logic to hint what was taking place. Typically I’d ask AI to re-explain the code line by line, which frequently revealed refined logical errors or edge circumstances I hadn’t thought of.
Importantly, this wasn’t nearly fixing bugs, it was additionally about studying. My Python expertise grew immensely by way of this course of, as I compelled myself to essentially perceive each line the AI generated. If I didn’t, I’d inevitably pay the worth later when a small tweak broke one thing downstream.
Today, I lean on an AI-powered built-in improvement atmosphere (IDE) to speed up debugging. However the precept is unchanged: don’t skip instrumentation and verification. If the AI can’t debug for you, fall again on print statements and your personal means to hint the issue to its supply. And at all times double verify any AI-generated code.
AI as an implementer, not inventor
This venture taught me that AI is implausible at taking a well-defined thought and turning it into working code. However if you happen to ask it to design an structure or invent a migration technique from scratch, it’ll most likely allow you to down. My most efficient workflow was to (1) design the method myself, (2) describe it intimately, (3) let the AI deal with the implementation and boilerplate, and (4) assessment, check, and refine the AI output.
Model management
I can’t stress sufficient the significance of model management, even for easy scripts. Each time I added a characteristic or fastened a bug, I made a commit. When a bug appeared days later, I might stroll again by way of my historical past and pinpoint the place issues broke. Positive, that is fundamental software program engineering, however once you’re working with AI, it’s much more crucial. The speed of change will increase, and your personal reminiscence of every modification is inevitably much less exhaustive.
The web impact of those practices was velocity with out chaos. We delivered far sooner than we might have in any other case, and the standard of the output considerably diminished post-migration cleanup.
The place AI fell brief
As worthwhile as AI was, it had many shortcomings. The cracks began to point out because the scripts grew in dimension and complexity:
- Context limits: When scripts bought longer, the AI misplaced observe of earlier code sections. It might add new standalone options, however integrating new logic into current, interdependent code? That usually failed except I spelled out precisely the place and methods to make modifications. I ought to be aware that right this moment’s newer fashions with bigger context home windows would possibly scale back a number of the points I bumped into with the migration scripts. However I believe that it’s nonetheless essential to be as particular as potential about what sections have to be up to date and with what logic.
- Failure to discover a working implementation: I discovered that generally the AI merely couldn’t resolve the issue as outlined within the immediate. If I requested for a change and it failed three or 4 instances, that was normally a sign to step again and check out one thing totally different – whether or not that meant prompting for an alternate method or writing the code myself.
- System understanding: Sure bugs or edge circumstances required a strong understanding of our methods, like how the CCMS handles ID values, or how competing case sensitivity guidelines throughout methods might journey issues up. This can be a essential space the place AI couldn’t assist me.
What I’d do in another way subsequent time
Right here’s my recommendation, if I needed to do it over again:
- Plan core libraries and conventions early: Resolve in your stack, naming schemes, and file construction on the outset and embrace them in each immediate. Inconsistencies right here led to time wasted refactoring scripts midstream. That stated, working in an editor-based instrument that’s conscious of your whole pipeline will assist to maintain your libraries constant from the outset.
- Sanitize every little thing: File names, IDs, casing, and different seemingly minor particulars could cause main downstream issues. Embody this steerage in your prompting boilerplate.
- Account for customized content material: Don’t assume all docs comply with the identical patterns and positively don’t assume the AI understands the nuances of your content material. Discover out early the place the outliers are. This upfront work will prevent time in the long term.
- Doc the advanced stuff: For any logic that takes various minutes to grasp, write down an intensive rationalization you possibly can refer again to later. There have been instances I needed to re-analyze sophisticated components of the scripts weeks later, when an in depth be aware would have set me again heading in the right direction.
One non-AI tip: maintain copies of your supply and transformed markup in a repository even after importing the transformed content material to your manufacturing tooling. I promise that you just’ll have to refer again to them.
AI as a associate, not a alternative
Reflecting on the venture, I can emphatically say that AI didn’t substitute my crucial considering. As a substitute, it amplified my expertise, serving to me work at a velocity and scale that might have been tough to attain alone, whereas streamlining the post-migration cleanup. However anytime I leaned too closely on AI with out cautious planning, I wasted time and needed to backtrack.
The true worth got here from pairing my area information and important considering with AI’s means to iterate rapidly and implement. Used thoughtfully, AI helped me ship a venture that grew to become a profession milestone.
In case you’re dealing with your personal daunting migration, or simply need to get extra out of AI in your workflow, I hope these classes prevent some ache, and possibly even encourage you to tackle a problem you might need thought was too large to deal with.
Discover extra tales on our Innovation channel and subscribe right here!